Patel Katie, Kay Richard, Rowell Lucy
Roche Global Development, Welwyn Garden City, UK.
Pharm Stat. 2006 Jul-Sep;5(3):213-24. doi: 10.1002/pst.213.
The proportional hazards (PH) model is routinely employed for the analysis of time-to-event data in medical research when it is required to assess the effect of an intervention in the presence of covariates. The assumption of PH required for the PH approach may not hold, especially in circumstances where the effect of the intervention is to delay or accelerate the onset of an event rather than to reduce or increase the overall proportion of subjects who observe the event through time. If the assumption of PH is violated, the results from a PH model will be difficult to generalize to situations where the length of follow-up is different to that used in the analysis. It is also difficult to translate the results into the effect upon the expected median duration of illness for a patient in a clinical setting. The accelerated failure time (AFT) approach is an alternative strategy for the analysis of time-to-event data and can be suitable even when hazards are not proportional and this family of models contains a certain form of PH as a special case. The framework can allow for different forms of the hazard function and may provide a closer description of the data in certain circumstances. In addition, the results of the AFT model may be easier to interpret and more relevant to clinicians, as they can be directly translated into expected reduction or prolongation of the median time to event, unlike the hazard ratio. We recommend that consideration is given to an AFT modelling approach in the analysis of time-to-event data in medical research.
当需要在存在协变量的情况下评估干预效果时，比例风险（PH）模型通常用于医学研究中对事件发生时间数据的分析。PH方法所需的PH假设可能不成立，特别是在干预效果是延迟或加速事件发生，而不是随着时间推移减少或增加观察到该事件的受试者总体比例的情况下。如果违反了PH假设，PH模型的结果将难以推广到随访时间与分析中使用的时间不同的情况。也很难将结果转化为对临床环境中患者预期疾病持续时间中位数的影响。加速失效时间（AFT）方法是分析事件发生时间数据的另一种策略，即使风险不成比例也可能适用，并且这类模型包含某种形式的PH作为特殊情况。该框架可以考虑不同形式的风险函数，并且在某些情况下可能对数据提供更贴近的描述。此外，AFT模型的结果可能更容易解释，并且与临床医生更相关，因为它们可以直接转化为事件发生中位数时间的预期缩短或延长，这与风险比不同。我们建议在医学研究中对事件发生时间数据的分析中考虑采用AFT建模方法。
